Category-based updating

Given a prior distribution over a finite outcome space, how is the distribution updated when one outcome is excluded (i.e., assigned probability 0)? We describe two experiments in which estimated probabilities seem to "stick" to salient events. The probabilities of such events remain relatively invariant through updating. Our results suggest that the credence assigned to a salient category is sometimes more basic than the credence assigned to the constituents that comprise the category.